acceptodds
Under review as a conference paper at ICLR 2027

GeoGSplat: Geometry-Grounded Non-Pixel-Aligned Gaussian Splatting

18 views

Abstract

Feed-forward Gaussian Splatting enables efficient novel view synthesis, but pixel-aligned prediction makes the primitive count grow with image resolution and view count, leading to substantial redundancy and increased computational cost. Non-pixel-aligned methods address this limitation by predicting a fixed number of Gaussians. Without explicit ray and depth constraints, these methods yield misplaced Gaussian centers, causing rendering artifacts under large viewpoint changes. To address this issue, we introduce **GeoGSplat**, a geometry-grounded feed-forward model that explicitly incorporates geometric priors into the non-pixel-aligned framework. A dual-branch encoder constructs *Anchors* and *Triplane* to capture coarse 3D structure and fine-grained scene features, respectively. Subsequently, a hierarchical decoder progressively decodes dense 3D points and Gaussian primitives, enabling 3D reconstruction and novel view synthesis. GeoGSplat surpasses both state-of-the-art pixel-aligned and non-pixel-aligned Gaussian Splatting methods on DTU and NRGBD in terms of novel view synthesis and reconstruction. The advantage of our geometry-grounded paradigm further increases as target viewpoints move farther from the input cameras, demonstrating stronger generalization to large viewpoint changes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, Feed-Forward 3DGS

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Qp4vOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GeoGSplat, a geometry-grounded feed-forward 3D Gaussian Splatting (3DGS) framework that operates in a non-pixel-aligned paradigm. While existing pixel-aligned methods suffer from a primitive count that scales linearly with input resolution and view count (N=NI×H×WN = N_I \times H \times W), recent non-pixel-aligned methods predict a fixed number of Gaussians via learnable tokens but discard explicit ray/depth geometric constraints, leading to poorly placed Gaussian centers and severe rendering artifacts under large viewpoint changes.

Critical strengths

The core diagnosis of why non-pixel-aligned 3DGS methods (like TokenGS) fail under large viewpoint changes—namely, photometric overfitting due to underconstrained Gaussian centers—is insightful and empirically well-substantiated (especially in Section 4.3, Figures 5 & 6, and Appendix A.1). By coupling fixed-budget Anchor queries with continuous Triplane feature fields, GeoGSplat maintains a constant primitive count (65K65\text{K}) regardless of whether 2, 4, or 12 views are input, while achieving superior geometric accuracy (Chamfer Distance reduced from 0.2770.277 to 0.0170.017 on NRGBD compared to TokenGS) and significantly better extrapolation to distant target views.

A common pitfall in large-scale feed-forward 3D vision papers is confounding architectural gains with dataset scaling or supervision differences. The authors proactively address this in Appendix A.1 (Table 4) by retraining TokenGS under the exact same five-dataset mixture, resolution, and explicit Chamfer Distance geometry loss (TokenGS-R + Geometry loss). Showing that GeoGSplat still outperforms TokenGS-R + Geometry loss by +6.07 dB+6.07\text{ dB} PSNR on DTU and reduces CD from 0.0440.044 to 0.0180.018 convincingly proves that the hierarchical Anchor-Triplane architecture itself—not merely the addition of GT 3D supervision—drives the performance leap.

Critical weaknesses

Joint self/cross-attention over M=4096M=4096 Anchor queries and nT=3×642=12,288n_T = 3 \times 64^2 = 12,288 Triplane queries across 12 shared attention blocks. Two stages of Triplane upsampling up to 3×256×2563 \times 256 \times 256 with 3D-to-2D cross-plane fusion and image-conditioned cross-attention at high spatial resolutions (Eq. 18). Computing symmetric Chamfer Distance and Farthest Point Sampling (FPS) on 65,53665,536 points during training (requiring 128 NVIDIA L20 GPUs). And The paper only reports the number of Gaussians (NUM) in Table 1, completely omitting feed-forward inference latency (ms), peak GPU memory usage, and rendering FPS compared to MVSplat, DepthSplat, AnySplat, and TokenGS. Without wall-clock runtime and memory metrics, the claim of "computational efficiency" remains only half-proven (rendering is fast, but encoding/decoding might be a bottleneck).

Unlike purely photometric feed-forward 3DGS methods (e.g., pixelSplat, MVSplat, TokenGS) that can be trained on posed RGB videos alone (such as RealEstate10K or ACID), GeoGSplat collapses without explicit ground-truth 3D point cloud supervision (w/o Geometry Losses in Table 3 drops PSNR from 19.7919.79 to 13.60 dB13.60\text{ dB} on DTU and CD degrades from 0.0190.019 to 0.2180.218). This restricts its training exclusively to datasets with accurate RGB-D or MVS depth. Training Only on Input Views: In Section 3.4 and 4.1, the authors note that during training, Lrender\mathcal{L}_{render} is computed only by rendering at the 4 input camera views without sampling held-out novel target views. In 3DGS, supervising only on context views notoriously encourages view-dependent overfitting or degenerate flat Gaussians. Although Lgeo\mathcal{L}_{geo} regularizes the 3D positions, it is unclear why standard novel-view photometric supervision was omitted and how much performance is lost/gained by this choice.

Questions for authors

Could you provide a quantitative breakdown of: (a) encoder-decoder feed-forward inference time (ms), (b) rasterization time (ms), and (c) peak GPU memory consumption across 2, 4, and 12 input views compared to MVSplat, DepthSplat, AnySplat, and TokenGS? Specifically, how expensive is the image-conditioned cross-attention (Eq. 18) at the 1282128^2 and 2562256^2 Triplane resolutions?

Equations (4) and (7) bound the Anchor centers and Triplane queries to [−1,1]3[-1, 1]^3 via tanh⁡(⋅)\tanh(\cdot). In Appendix B.5, you mention that during geometric evaluation, "the ground-truth point cloud determines its scale" to normalize the scene into [−1,1]3[-1, 1]^3. During inference (before GT geometry is known), how is the scene coordinate frame and scale defined solely from the input camera poses πi\pi_i so that unbounded scenes (like Mip-NeRF 360 in Appendix A.3) fit properly inside the [−1,1]3[-1, 1]^3 Triplane bounds without clipping background geometry?

Why did you choose to compute Lrender\mathcal{L}_{render} only on the 4 input views during training rather than held-out target views (which is standard in feed-forward NVS)? Does supervising on novel target views further improve PSNR/LPIPS, or allow the model to rely less on GT depth supervision (Lgeo\mathcal{L}_{geo})?

Notation in Eq. (8): In Eq. (8), ΔXi=CrossAtt(Qioff,Figeo,Figeo)\Delta X_i = \text{CrossAtt}(Q_i^{off}, F_i^{geo}, F_i^{geo}). Are the learnable offset tokens Qoff∈RK×CQ^{off} \in \mathbb{R}^{K \times C} shared across all MM Anchors, or distinct per Anchor? Moreover, since Figeo∈RK×3CTF_i^{geo} \in \mathbb{R}^{K \times 3C_T} already has a 1-to-1 correspondence with the KK sampled local reference points Xˉi\bar{X}_i, why use an all-to-all K×KK \times K cross-attention with learnable queries instead of a point-wise MLP or local self-attention directly on FigeoF_i^{geo}

3 comments
AuthorsOpenReview ↗

Official Response by Authors (1/3)

We thank the reviewer for the constructive feedback. We address the main weaknesses and their related questions below.

W1: Computational efficiency (Q1)

Summary: The efficiency emphasized in our paper is the representation efficiency of non-pixel-aligned methods over pixel-aligned methods. Our additional experiments demonstrate that this representation efficiency translates into substantial practical computational gains while maintaining superior NVS and reconstruction quality, without significant overhead from high-resolution triplane cross-attention.

Computational profiling. GeoGSplat's representation efficiency translates into substantial practical computational gains while achieving superior NVS and reconstruction quality. For feed-forward 3DGS, practical efficiency is primarily reflected in reconstruction speed and rendering throughput. We therefore profile GeoGSplat (351.9M parameters) against AnySplat across 12–64 input views, reporting inference latency, FLOPs, rendering throughput, and peak GPU memory. The computational advantages become more pronounced as the number of input views increases.

ViewsMethodGaussians ↓Inference (ms) ↓FLOPs (TF) ↓Rendering FPS ↑Peak memory (GiB) ↓
12GeoGSplat65,536 (4.0%)299 (71.7%)38.79 (84.5%)493 (1.41×)3.99 (65.7%)
AnySplat1,651,46041745.903506.07
24GeoGSplat65,536 (2.1%)443 (45.6%)66.07 (54.3%)619 (2.47×)4.14 (44.1%)
AnySplat3,164,598972121.782519.38
48GeoGSplat65,536 (1.1%)856 (34.3%)150.31 (41.4%)661 (4.11×)4.44 (27.8%)
AnySplat5,742,4202,494363.4716115.99
64GeoGSplat65,536 (0.9%)1,229 (32.4%)228.46 (38.6%)578 (4.22×)4.64 (22.8%)
AnySplat7,074,1483,791591.2113720.38

Parentheses indicate GeoGSplat's resource usage as a percentage of AnySplat's for Gaussians, inference latency, FLOPs, and peak GPU memory, and relative rendering throughput (×) for FPS.

High-resolution triplane cross-attention. The two high-resolution triplane cross-attention stages are lightweight. We pool image key/value features to 64 tokens before cross-attention. We further measured the latency of the two high-resolution triplane refinement stages.

Views128² cross-attn256² cross-attnTotal latencyNetwork ratio
120.559 ms0.937 ms1.496 ms0.50%
240.592 ms0.948 ms1.540 ms0.35%
480.602 ms0.974 ms1.576 ms0.18%
640.571 ms0.957 ms1.528 ms0.12%

AuthorsOpenReview ↗

Official Response by Authors (2/3)

W2: Dependence on geometry supervision and input-view training (Q3)

Summary: GeoGSplat's improvements are enabled by our architecture, which is closely integrated with geometry supervision, rather than by geometry supervision alone. Simply adding geometry loss without a design that effectively leverages geometry is insufficient. Although geometry supervision is important to our method, GeoGSplat maintains stable NVS performance even with imperfect depth annotations. Moreover, additional novel-view photometric supervision does not yield significant improvements over input-view-only supervision.

Dependence on geometry supervision. As demonstrated by our additional TokenGS retraining experiments in the General Response, simply adding geometry loss is insufficient to match GeoGSplat. Even with its original novel-view photometric supervision, visibility loss, and additional geometry loss, TokenGS still underperforms GeoGSplat, which achieves 52.6% lower CD on DTU and 41.4% lower CD on NRGBD.

Unlike TokenGS, which directly predicts Gaussians from tokens, GeoGSplat explicitly models intermediate geometry through Anchors and dense points, complemented by a Triplane representation that encodes fine-grained high-dimensional scene features queryable at arbitrary 3D coordinates. This design enables more effective use of geometry, while the geometry supervision makes these intermediate structures geometrically meaningful.

We acknowledge that geometry supervision is important for our explicit intermediate geometry. However, GeoGSplat does not require perfectly accurate or complete GT depth. Experiments with noisy and missing depth using our lightweight model show PSNR changes of at most 0.05 dB, demonstrating stable NVS performance despite some degradation in CD under noisy depth.

Training settingDTU PSNR ↑DTU CD ↓NRGBD PSNR ↑NRGBD CD ↓
Original19.790.01923.700.019
Noisy depth19.740.03023.750.025
Missing depth19.760.01923.670.020

Furthermore, Tables 3 and 4 support the effectiveness of our design. Component ablations consistently degrade performance, while the controlled comparison in Table 4 shows higher PSNR and lower CD under matched training conditions and geometry supervision.

Input-view-only photometric supervision. Novel-view photometric supervision alone cannot resolve all geometric ambiguities. As shown in Fig. 6, TokenGS still produces geometric outliers, even after opacity filtering, while points unprojected from rendered depth remain displaced from GT geometry. These errors may have limited impact near input views but become increasingly apparent under large viewpoint changes (Fig. 5).

Moreover, we also tested novel-view photometric supervision on GeoGSplat, but did not observe a significant improvement over input-view-only supervision. Similarly, AnySplat adopts context-view-only photometric supervision.

AuthorsOpenReview ↗

Official Response by Authors (3/3)

Q2: Scene normalization and unbounded scenes

Summary: GeoGSplat does not require GT geometry for inference-time scene normalization and maintains stable NVS performance with approximate scene scales. It also achieves strong NVS and geometric reconstruction performance on large-scale scenes without scene contraction or background masking.

Inference-time normalization. GeoGSplat does not require GT geometry to estimate scene normalization parameters at inference time. The scene center can be estimated from input camera poses using either camera-ray intersection or mean camera centering, depending on the dataset configuration. Similar to TokenGS, GeoGSplat then normalizes camera translations using a dataset-dependent fixed scale factor.

To verify robustness to scene-scale estimation, we additionally evaluated GeoGSplat on DTU using a fixed scale factor independent of GT geometry. GeoGSplat still outperforms all baselines in Table 1 without GT scale, even though these baselines are evaluated with GT scale.

Normalization strategyPSNR ↑SSIM ↑LPIPS ↓
GT scale20.480.7710.137
Fixed scale (w/o GT geometry)20.350.7590.148

Unbounded scenes. GeoGSplat achieves the highest PSNR (21.59 dB) and lowest CD (0.015) on Mip-NeRF360 without scene contraction or background masking. For distant backgrounds such as skies, finite Gaussians approximate their appearance rather than physically accurate infinite-depth geometry. Nevertheless, these results demonstrate that GeoGSplat remains effective on large-scale scenes with unbounded backgrounds.


Q4: Anchor-to-Point cross-attention

Summary: Local cross-attention enables more effective aggregation of geometric features within each anchor, improving both NVS quality and geometric reconstruction compared with pointwise MLP prediction.

Offset queries. The same set of K = 16 learnable offset queries is shared across all anchors.

Cross-attention design. Cross-attention enables each offset query to adaptively aggregate local Triplane features, allowing point offsets to be predicted using the overall local geometry rather than individual point features. To validate its effectiveness, we replaced cross-attention with a pointwise MLP, resulting in a 0.30 dB PSNR decrease and a 0.002 increase in CD.

Reviewer_HHDwOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of misplaced Gaussian centers in non-pixel-aligned feed-forward Gaussian Splatting, which lacks explicit geometric constraints and therefore produces rendering artifacts under large viewpoint changes. The proposed GeoGSplat consists of (1) a dual-branch encoder that constructs Anchors with learnable centers and adaptive radii, together with a Triplane that encodes fine-grained scene features, and (2) a hierarchical decoder that recovers dense points through Anchor-to-Point decoding and predicts Gaussians through Point-to-Gaussian decoding, with a fixed budget of 65K Gaussians. The model is trained with Chamfer losses on point clouds constructed from ground-truth depth, in addition to photometric losses. Experiments on DTU and NRGBD show improved novel view synthesis and 3D reconstruction over both pixel-aligned and non-pixel-aligned methods, and the advantage over TokenGS increases as the target-context distance grows.

Critical strengths
  1. The method achieves substantial improvements in both NVS and geometry with a small, fixed number of Gaussians. Compared with TokenGS under the same 65K Gaussian budget, GeoGSplat improves DTU PSNR from 17.71 to 20.48 dB and LPIPS from 0.354 to 0.137 (Table 1), and reduces NRGBD CD from 0.277 to 0.017 (Table 2). It also outperforms AnySplat∗, which uses approximately 10× more Gaussians (624K) on NRGBD, in both NVS and CD. These results support the practical value of geometric grounding for efficient feed-forward 3DGS.
  2. The paper concretely analyzes how geometric errors lead to NVS degradation. Fig. 5 shows that the gap over TokenGS widens as the viewpoint change increases (up to approximately 3.2 dB), and Fig. 6 shows that the outlier centers of TokenGS remain even after opacity filtering. This analysis convincingly supports the validity of the problem formulation.
Critical weaknesses
  1. The paper does not isolate whether the improvements come from the proposed architecture or from ground-truth 3D supervision and different training data. GeoGSplat is trained with Chamfer losses on point clouds constructed from ground-truth depth (Sec. 3.4, Eq. 10). In contrast, TokenGS in the main tables is a publicly released model trained on different data (App. A.1), and the checkpoint sources and training data of the other baselines are not specified. TokenGS is the only baseline retrained under the same conditions. On NRGBD, adding only the geometry loss to the retrained TokenGS (TokenGS-R) improves PSNR from 22.48 to 24.17 dB and CD from 0.180 to 0.029 (Table 4). Thus, 1.69 dB of the 2.30 dB PSNR gap to GeoGSplat (24.78 dB PSNR / 0.017 CD) is closed by the geometry loss alone, which is inconsistent with the conclusion in App. A.1 that the gains come from the geometry-grounded architecture itself. On DTU, a large gap remains, but TokenGS-R (13.75 dB) is approximately 4 dB below the public TokenGS (17.71 dB). GeoGSplat is trained by rendering only at the input cameras without sampling target views (Sec. 4.1), and TokenGS-R is described only as trained under “the same setting” as GeoGSplat (App. A.1). Since the original TokenGS uses separate supervision views and a visibility loss, the retraining setting may have disadvantaged TokenGS. Furthermore, GeoGSplat without any geometry losses (DTU 13.60 dB / NRGBD 18.53 dB, Table 3) performs worse than TokenGS-R (13.75 dB / 22.48 dB). Even considering that Table 3 uses a smaller model (Table 8), the architectural advantage does not emerge without 3D supervision. Notably, Sec. 4.4 itself attributes this drop to “the importance of geometric supervision for maintaining reliable geometric priors,” which is in tension with the claim in App. A.1 that the gains “do not simply come from adding training data or geometric supervision.” In terms of novelty, AnchorSplat [CVPR'26], which predicts anchor-aligned Gaussians guided by 3D geometric priors independently of image resolution and the number of views, appears only once in a citation list in Sec. 2, without comparison or discussion of the differences. Since the central claim of the paper is that the geometry-grounded architecture itself drives the performance (Sec. 2 and the contributions), this issue directly affects my assessment.
  2. The claims of efficiency and scalability are insufficiently supported. The paper repeatedly emphasizes efficiency (Abstract, contributions, and Sec. 3.1), but reports only Gaussian counts (Tables 1 and 5), without comparisons of inference time, peak memory, FLOPs, or parameter counts. Training requires 320K iterations on 128 NVIDIA L20 GPUs, but the training time is not reported. The number of Gaussians is fixed at M × K = 4096 × 16, and the Anchor centers are bounded within the normalized volume by tanh (Eq. 4). However, on Mip-NeRF 360, the PSNR improvement over TokenGS is only 0.08 dB (21.51 → 21.59 dB, Table 6), which provides only weak support for the claim that the method generalizes well to large-scale scenes (App. A.3). Without an analysis of how quality and cost change with M and K, it is difficult to assess scalability with respect to scene size.
Questions for authors
  1. Source of improvements [Clarification / Additional experiments]: Was the rendering supervision for TokenGS-R applied only at the input views, and was the original visibility loss of TokenGS included? Can the authors provide results of TokenGS retrained with its original recipe (rendering at separate supervision views with the visibility loss), as well as a comparison with a main-scale GeoGSplat trained without geometry losses?

  2. Baseline training conditions [Clarification]: What are the checkpoint sources, training data, and use of depth/3D supervision for the baselines other than TokenGS (MVSplat, DepthSplat, AnySplat, and VolSplat)? Can at least one pixel-aligned method (e.g., DepthSplat) be retrained on the same data for comparison?

  3. Relation to AnchorSplat [Clarification]: How does GeoGSplat differ in design from AnchorSplat [CVPR'26], and can the authors provide a quantitative comparison?

  4. Scene normalization at test time [Clarification]: How are the model inputs (cameras and scene) mapped into the normalized space during training and inference, and what determines the scale? App. B.5 sets the scale of the evaluation coordinate system using the ground-truth point cloud. Is this information used as input to GeoGSplat at inference, or applied differently to the baselines? This directly affects the fairness of the CD comparisons.

  5. Efficiency [Additional experiments]: Can the authors report inference time, peak memory, FLOPs, and parameter counts compared with the baselines for 2, 4, and 12 input views, together with the total training time (GPU-hours) measured in the same environment for fair comparisons?

  6. Scalability and design choices [Additional experiments]: How do quality and cost change with M and K, particularly on larger scenes such as Mip-NeRF 360? Can the authors provide ablations on adaptive versus fixed radii (Eq. 5), the values of γ_min and γ_max, fixed random reference points versus grid- or FPS-based initialization (Sec. 3.2, Sec. 3.3, and App. B.3), and the 15K-iteration geometry-only warm-up (Sec. 3.4)?

  7. Generality of the viewpoint-change analysis [Clarification / Additional experiments]: Fig. 5 is presented as evidence that the advantage over TokenGS increases with the target-context distance. However, the evaluation conditions are limited in two respects.

First, PSNR is computed only on pixels visible in at least one context view (App. B.6). As viewpoints move farther, unobserved regions grow, and these are the regions where the strengths and weaknesses of the two methods are most likely to differ. GeoGSplat is supervised by point clouds constructed from input-view depth and may leave holes in these regions, whereas TokenGS claims scene completion beyond the input views as an advantage. Excluding these regions may therefore mask the actual difference. Moreover, only PSNR is reported, so changes in perceptual quality remain unknown.

Second, TokenGS, which lacks geometric priors, is the only method compared. Since the paper argues that geometric priors provide robustness to large viewpoint changes, comparisons with pixel-aligned methods that rely on per-pixel depth priors (AnySplat∗ and DepthSplat) are necessary to distinguish whether this robustness comes from the design of GeoGSplat or simply from the presence of geometric priors. Could the authors provide (a) full-image PSNR and LPIPS, including unobserved regions, as functions of the target-context distance, and (b) distance-wise PSNR (and LPIPS) of AnySplat∗ and DepthSplat along the same target-view trajectories, added to Fig. 5? These results would clarify whether the claim of “stronger generalization to large viewpoint changes” in the Abstract is supported.

  1. Reproducibility and limitations [Clarification]: What are the values of λ_s and λ_p (Eq. 11) and of γ_min and γ_max (Eq. 5), and do the authors plan to release the code? A discussion of limitations and failure cases (e.g., thin structures and large-scale scenes) should also be added.
4 comments
AuthorsOpenReview ↗

Official Response by Authors (1/4)

We thank the reviewer for the constructive feedback. We address the main weaknesses and their related questions below.

W1: Source of improvements and dependence on geometry supervision (Q1)

Summary: GeoGSplat's improvements are enabled by our architecture, which is closely integrated with geometry supervision, rather than by geometry supervision alone. Simply adding geometry loss without a design that effectively leverages geometry is insufficient. Although geometry supervision is important to our method, GeoGSplat maintains stable NVS performance even with imperfect depth annotations.

Simply adding geometry loss is insufficient. As requested, we retrained TokenGS using its original novel-view photometric supervision and visibility loss, both with and without additional geometry loss. Although geometry loss improves TokenGS, it still substantially underperforms GeoGSplat, particularly in reconstruction.

MethodDTU PSNR ↑DTU SSIM ↑DTU LPIPS ↓DTU CD ↓NRGBD PSNR ↑NRGBD SSIM ↑NRGBD LPIPS ↓NRGBD CD ↓
TokenGS15.240.5930.3170.46023.240.7780.1760.191
TokenGS + geometry loss16.370.6080.3140.03824.430.8070.1530.029
GeoGSplat20.480.7710.1370.01824.780.8470.1200.017

Unlike TokenGS, which directly predicts Gaussians from tokens, GeoGSplat explicitly models intermediate geometry through Anchors and dense points, complemented by a Triplane representation that encodes fine-grained high-dimensional scene features queryable at arbitrary 3D coordinates. This design enables more effective use of geometry, while the geometry supervision makes these intermediate structures geometrically meaningful. Therefore, the performance degradation observed when removing geometry supervision is consistent with our architecture.

Robustness to imperfect depth. Although geometry supervision is important for learning our explicit intermediate geometry, GeoGSplat does not require perfectly accurate or complete GT depth. We conducted additional experiments using our lightweight model. Specifically, we add Gaussian noise (σ = 5%) to randomly selected 10%–90% of depth pixels or remove depth supervision from one of the four input views. The results demonstrate that GeoGSplat maintains stable NVS performance under both settings.

Training settingDTU PSNR ↑DTU CD ↓NRGBD PSNR ↑NRGBD CD ↓
Original19.790.01923.700.019
Noisy depth19.740.03023.750.025
Missing depth19.760.01923.670.020

Further architectural evidence. Tables 3 and 4 further support the effectiveness of our design. Component ablations in Table 3 consistently degrade performance, while the controlled comparison in Table 4 shows higher PSNR and lower CD under matched training conditions and geometry supervision. These results further demonstrate that our architecture enables more effective use of geometry supervision, rather than merely benefiting from an additional geometry loss.

AuthorsOpenReview ↗

Official Response by Authors (2/4)

W2: Computational Efficiency (Q5)

Summary: The efficiency emphasized in our paper is the representation efficiency of non-pixel-aligned methods over pixel-aligned methods. Our additional experiments demonstrate that this representation efficiency translates into substantial practical computational gains while maintaining superior NVS and reconstruction quality.

Computational profiling. GeoGSplat's representation efficiency translates into substantial practical computational gains while achieving superior NVS and reconstruction quality. For feed-forward 3DGS, practical efficiency is primarily reflected in reconstruction speed and rendering throughput. We therefore profile GeoGSplat (351.9M parameters) against AnySplat across 12–64 input views, reporting inference latency, FLOPs, rendering throughput, and peak GPU memory. The computational advantages become more pronounced as the number of input views increases.

ViewsMethodGaussians ↓Inference (ms) ↓FLOPs (TF) ↓Rendering FPS ↑Peak memory (GiB) ↓
12GeoGSplat65,536 (4.0%)299 (71.7%)38.79 (84.5%)493 (1.41×)3.99 (65.7%)
AnySplat1,651,46041745.903506.07
24GeoGSplat65,536 (2.1%)443 (45.6%)66.07 (54.3%)619 (2.47×)4.14 (44.1%)
AnySplat3,164,598972121.782519.38
48GeoGSplat65,536 (1.1%)856 (34.3%)150.31 (41.4%)661 (4.11×)4.44 (27.8%)
AnySplat5,742,4202,494363.4716115.99
64GeoGSplat65,536 (0.9%)1,229 (32.4%)228.46 (38.6%)578 (4.22×)4.64 (22.8%)
AnySplat7,074,1483,791591.2113720.38

Parentheses indicate GeoGSplat's resource usage as a percentage of AnySplat's for Gaussians, inference latency, FLOPs, and peak GPU memory, and relative rendering throughput (×) for FPS.

AuthorsOpenReview ↗

Official Response by Authors (3/4)

Q2: Baseline Training Conditions

Summary: We use the officially released pretrained checkpoints for the selected baselines, which already benefit from substantial pretrained geometric priors. Despite these advantages, GeoGSplat achieves superior NVS performance, particularly compared with AnySplat, which uses broader training data and additional geometry supervision.

Baseline training conditions. The training datasets, pretrained geometric priors, and supervision strategies of the baselines are summarized below.

MethodTraining dataGeometric prior / supervision
MVSplatRealEstate10KPretrained UniMatch; RGB rendering loss
DepthSplatRealEstate10KPretrained Depth Anything V2 + UniMatch; RGB rendering loss
VolSplatRealEstate10KPretrained Depth Anything V2 + UniMatch; RGB rendering loss
AnySplatNine datasets (including our five training datasets)Pretrained VGGT; RGB rendering loss, depth/pose distillation, and rendered-depth consistency loss

Comparison with AnySplat. AnySplat benefits from pretrained VGGT and additional geometry supervision, while being trained on nine datasets, including all five used by GeoGSplat. Nevertheless, under the same GT-pose evaluation protocol, GeoGSplat outperforms AnySplat* by 2.94 dB on DTU and 4.26 dB on NRGBD, demonstrating its effectiveness despite using less extensive training data and geometric priors.

DatasetAnySplat* PSNR ↑GeoGSplat PSNR ↑
DTU17.5420.48 (+2.94)
NRGBD20.5224.78 (+4.26)

Q3: Comparison with AnchorSplat

Summary: GeoGSplat differs from AnchorSplat in its anchor representation, hierarchical decoding, and geometry supervision strategy.

Differences from AnchorSplat. First, AnchorSplat constructs point-based anchors by unprojecting depth predicted by pretrained MapAnything and applying FPS. In contrast, GeoGSplat constructs sphere-based Anchors with learnable centers and adaptive radii.

Second, AnchorSplat directly predicts Gaussians around its anchors and employs rendered-depth loss for geometry supervision. In contrast, GeoGSplat adopts Anchor-to-Point and Point-to-Gaussian decoding to progressively reconstruct dense geometry, leveraging Triplane features queryable at arbitrary 3D coordinates. Moreover, we directly supervise both anchor centers and dense points.

Quantitative comparison. Since AnchorSplat does not provide pretrained model weights, we cannot conduct a quantitative comparison.


Q4: Scene Normalization at Test Time

Summary: GeoGSplat does not require GT scene scale at inference time, and its NVS performance remains stable with approximate scene-scale normalization. All predicted geometries are transformed into the same coordinate system for fair Chamfer distance evaluation.

Scene normalization. We clarify that GT geometry is used for scene-scale calibration only in our standard evaluation, where the same procedure is applied to all baselines to ensure fair comparisons. At inference time, GeoGSplat requires only posed RGB images and approximate scene-scale normalization, without accessing GT geometry. Similar scene-scale or depth-range normalization is commonly used by pose-conditioned feed-forward 3DGS methods, including DepthSplat, VolSplat, and TokenGS.

To verify robustness to scene-scale estimation, we additionally evaluated GeoGSplat on DTU using a fixed scale factor independent of GT geometry. GeoGSplat still outperforms all baselines in Table 1 without GT scale, even though these baselines are evaluated with GT-based scale calibration.

Normalization strategyPSNR ↑SSIM ↑LPIPS ↓
GT scale20.480.7710.137
Fixed scale (w/o GT geometry)20.350.7590.148

Chamfer distance evaluation. As described in Appendix B.5, all predicted geometries are transformed into the same GT coordinate system before computing Chamfer distance, ensuring fair and consistent geometric evaluation across methods.

AuthorsOpenReview ↗

Official Response by Authors (4/4)

Q6: Scalability to Larger Scenes and Gaussian Budget

Summary: GeoGSplat achieves accurate reconstruction on large-scale scenes with a compact Gaussian representation. Our additional ablation further demonstrates that increasing the Gaussian budget yields only marginal NVS improvements without improving geometric accuracy.

Large-scale scene reconstruction. On Mip-NeRF360, GeoGSplat achieves higher PSNR than TokenGS while substantially reducing CD from 0.132 to 0.015, demonstrating its effectiveness on large-scale scenes with a fixed Gaussian budget.

Gaussian budget ablation. To further investigate the effect of Gaussian count, we conducted an additional ablation using our lightweight model. With M fixed at 4096, increasing K from 16 to 64 quadruples the number of Gaussians. The results indicate that increasing the Gaussian budget does not necessarily improve both NVS quality and geometric accuracy.

MKGaussiansPSNR ↑SSIM ↑LPIPS ↓CD ↓
40961665,53620.750.5260.3300.016
409664262,14420.870.5320.3230.018

Other ablations. Due to limited computational resources, we have not yet completed the remaining ablation studies. We will conduct these experiments as soon as possible and report the results.


Q7: Generality of the Viewpoint-Change Analysis

Summary: GeoGSplat maintains its advantages under large viewpoint changes even under full-image evaluation, including unobserved regions. It achieves the highest full-image PSNR across all distances, demonstrating that its advantages do not depend on visibility masking.

Full-image evaluation. To address this concern, we conducted additional full-image PSNR and LPIPS evaluations against TokenGS and the geometry-guided pixel-aligned method AnySplat*. GeoGSplat consistently achieves the highest PSNR across all target-context distances, with its advantage over TokenGS generally increasing as viewpoint changes become larger. At the farthest distance, where 41.51% of pixels are unobserved, GeoGSplat still outperforms TokenGS and AnySplat* by 2.03 dB and 2.76 dB PSNR, respectively.

DistanceInvisible ratioGeoGSplat PSNR ↑TokenGS PSNR ↑AnySplat* PSNR ↑GeoGSplat LPIPS ↓TokenGS LPIPS ↓AnySplat* LPIPS ↓
0.000.00%25.3125.0823.540.1030.1310.106
0.112.58%24.6324.5022.340.1090.1400.125
0.227.98%22.6121.8420.650.1290.1830.153
0.3310.88%21.9721.0119.780.1390.1940.172
0.4413.99%21.6420.2819.200.1610.2170.192
0.5617.14%20.9819.6218.750.1770.2380.203
0.6719.27%20.3418.9318.440.2000.2620.202
0.7822.52%19.5117.8717.370.2240.2830.218
0.8925.03%18.5616.6515.890.2560.3240.248
1.0041.51%14.7312.7011.970.3890.4470.386

Fig. 5 excludes unobserved regions to isolate reconstruction quality. Feed-forward 3D Gaussian Splatting is a reconstruction framework rather than a generative model, aiming to accurately reconstruct regions supported by input observations rather than generate unobserved content.

Our main NVS results in Table 1 already use full-image evaluation. GeoGSplat consistently outperforms TokenGS and AnySplat* on both DTU and NRGBD, further confirming that its advantages do not depend on visibility masking.

DatasetTokenGS PSNR ↑AnySplat* PSNR ↑GeoGSplat PSNR ↑
DTU17.7117.5420.48
NRGBD23.8220.5224.78

Q8: Reproducibility and Limitations

Implementation details. The hyperparameters used in our experiments are λ_s = 0.4, λ_p = 0.1, γ_min = 0.375, and γ_max = 1.5.

Code release. We will release the code after acceptance.

Limitations. GeoGSplat employs a fixed Gaussian budget, which may lack the flexibility to adapt to scenes of varying complexity.

Reviewer_bBqhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GeoGSplat proposes a geometry-grounded non-pixel-aligned 3DGS framework for efficient novel view synthesis and 3D reconstruction. It introduces a dual-branch encoder that combines Anchors for coarse 3D structure and Triplane features for fine-grained scene information, followed y a hierarchical decoder to recover dense 3D points and Gaussian splats.

Critical strengths
  1. GeoGSplat addresses a meaning weakness/tradeoff of non-pixel-aligned GS mothods. Pixel-aligned methods are effective for rendering quality, but the Gaussian centers are unconstrained, causing geometric distortion or "background collapse" issue at novel viewpoints far away from training cameras.

  2. The experiment is well-targeted to support its claim. It reports consistent improvements on DTU and NRGBD, especially in geometric accuracy. The component ablations of retraining TokenGS indicate that the gains are not solely due to additional training data. These results support their claim of better geometric grounding produces better feed-foward 3DGS.

Critical weaknesses
  1. The improvements may rely heavily on the groundtruth depth-detrived pointclouds. The ablations show that performance deteriorates sharply when the geometry losses are removed. This makes it unclear how well the method would transfer to settings without accurate depth supervision and make comparisons with methods trained only from images difficult. Since geometric grounding is the central claimed contribution, its dependence on strong supervision naturally limits the scopr of that contribution.

  2. The efficiency and generalization claims are not fully supported by the evaluation. Although the output is fixed at 65K Gaussians, the paper reports training for 320k iterations on 128 NVIDIA L20 GPUs, without providing inference-time, peak memory usage,or any end to end computational comparisons. Also, the large viewpoint change experiment evaluates PSNR only on pixels visible in at least one input view, while the broader evaluation assumes groundtruth poses and geometry-based normalization. These choices make it difficult to determine whether the method is efficient and robust in realistic settings.

Questions for authors
  1. The geometry loss uses ground-truth point sets obtained by unprojecting valid depth from the input views. What depth source is used for each training dataset, such as sensor depth, rendered depth, or estimated depth, and how sensitive is performance to depth noise or missing depth? This would clarify whether the method’s gains can be expected in settings without high-quality geometric supervision.

  2. The paper motivates GeoGSplat partly through efficiency, but reports training on 128 NVIDIA L20 GPUs and does not provide matched computational measurements. Could the authors report parameter count, training time, inference latency, peak memory, and rendering throughput relative to the principal baselines? A fixed Gaussian count alone does not establish lower end-to-end cost.

  3. In the large-viewpoint experiment, PSNR is computed only over target pixels visible in at least one context view. How do the conclusions change when evaluating the complete target image, and what proportion of pixels is excluded at each target-context distance? This is important for determining whether the reported robustness extends to genuinely novel content or only to previously observed regions.

5 comments
AuthorsOpenReview ↗

Official Response by Authors (1/3)

We thank the reviewer for the constructive feedback. We address the main weaknesses and their related questions below.

W1: Scope of contribution and dependence on geometry supervision (Q1)

Summary: Geometry supervision is important for GeoGSplat, but it does not require perfectly accurate or complete depth annotations. Both the imperfect depth sources used in training and our additional robustness experiments support this conclusion.

Depth sources. Our training datasets contain geometric annotations from different sources, as summarized below. In particular, only 30% of CO3D videos with accurate camera tracking passed point-cloud quality checks. Despite these imperfect annotations, GeoGSplat achieves strong NVS and reconstruction performance on DTU and NRGBD, demonstrating its robustness to imperfect training geometry annotations.

DatasetDepth sourcePotential limitations
ScanNet++Depth rendered from laser-scanned meshesDepends on reconstructed mesh quality
ARKitScenesApple LiDARMeasurement noise and missing regions
WildRGB-DApple LiDARMeasurement noise and missing regions
BlendedMVSMulti-view reconstructionPotential reconstruction failures
CO3DMulti-view reconstructionPotential unreliable geometry

Robustness to noisy and missing depth. To further investigate the robustness to imperfect depth supervision, we conducted two additional ablation experiments using our lightweight model. Specifically, we add Gaussian noise (σ = 5%) to randomly selected 10%–90% of depth pixels or remove depth supervision from one of the four input views. The results demonstrate that GeoGSplat maintains stable NVS performance under both settings.

Training settingDTU PSNR ↑DTU CD ↓NRGBD PSNR ↑NRGBD CD ↓
Original19.790.01923.700.019
Noisy depth19.740.03023.750.025
Missing depth19.760.01923.670.020

Effective integration of geometry. Simply adding geometry loss to TokenGS is insufficient to match GeoGSplat, as shown in our new experiments in the General Response. Together with the component ablations in Table 3, these results demonstrate that our explicit intermediate geometry and architecture enable more effective use of geometry supervision, rather than merely benefiting from an additional geometry loss.

AuthorsOpenReview ↗

Official Response by Authors (2/3)

W2: Efficiency and robustness in realistic settings (Q2)

Summary: The efficiency emphasized in our paper is the representation efficiency of non-pixel-aligned methods over pixel-aligned methods. Our additional experiments demonstrate that this representation efficiency translates into substantial practical computational gains while maintaining superior NVS and reconstruction quality. Furthermore, GeoGSplat maintains stable NVS performance with approximate scene scales and estimated camera poses, supporting its practical applicability.

Computational profiling. GeoGSplat's representation efficiency translates into substantial practical computational gains while achieving superior NVS and reconstruction quality. For feed-forward 3DGS, practical efficiency is primarily reflected in reconstruction speed and rendering throughput. We therefore profile GeoGSplat (351.9M parameters) against AnySplat across 12–64 input views, reporting inference latency, FLOPs, rendering throughput, and peak GPU memory. The computational advantages become more pronounced as the number of input views increases.

ViewsMethodGaussians ↓Inference (ms) ↓FLOPs (TF) ↓Rendering FPS ↑Peak memory (GiB) ↓
12GeoGSplat65,536 (4.0%)299 (71.7%)38.79 (84.5%)493 (1.41×)3.99 (65.7%)
AnySplat1,651,46041745.903506.07
24GeoGSplat65,536 (2.1%)443 (45.6%)66.07 (54.3%)619 (2.47×)4.14 (44.1%)
AnySplat3,164,598972121.782519.38
48GeoGSplat65,536 (1.1%)856 (34.3%)150.31 (41.4%)661 (4.11×)4.44 (27.8%)
AnySplat5,742,4202,494363.4716115.99
64GeoGSplat65,536 (0.9%)1,229 (32.4%)228.46 (38.6%)578 (4.22×)4.64 (22.8%)
AnySplat7,074,1483,791591.2113720.38

Parentheses indicate GeoGSplat's resource usage as a percentage of AnySplat's for Gaussians, inference latency, FLOPs, and peak GPU memory, and relative rendering throughput (×) for FPS.

GT scene scale. GeoGSplat does not require GT scene scale at inference time. GT geometry is used for scene-scale normalization only in our standard evaluation, where the same calibration is applied to all baselines to ensure fair comparisons. In practice, GeoGSplat only requires posed RGB images and approximate scene-scale normalization. Similar normalization is also commonly used by pose-conditioned feed-forward 3DGS methods, including DepthSplat, VolSplat, and TokenGS.

To verify robustness to scene-scale estimation, we additionally evaluated GeoGSplat on DTU using a fixed scale factor independent of GT geometry. GeoGSplat still outperforms all baselines in Table 1 without GT scale, even though these baselines are evaluated with GT scale.

Normalization strategyPSNR ↑SSIM ↑LPIPS ↓
GT scale20.480.7710.137
Fixed scale (w/o GT geometry)20.350.7590.148

Estimated camera poses. GT camera poses are not necessary for GeoGSplat at inference time. Practical inference can directly use poses estimated by SfM, SLAM, or learned pose estimators. We additionally evaluated GeoGSplat using VGGT-Omega estimated poses on NRGBD, demonstrating stable NVS performance without GT camera poses.

Camera posePSNR ↑SSIM ↑LPIPS ↓
GT pose24.780.8470.120
VGGT-Omega pose24.440.8350.123
AuthorsOpenReview ↗

Official Response by Authors (3/3)

Q3: Generalization under large viewpoint changes

Summary: GeoGSplat maintains its advantages under large viewpoint changes even under full-image evaluation, including unobserved regions. It achieves the highest full-image PSNR across all distances, demonstrating that its advantages do not depend on visibility masking.

Full-image evaluation. To address this concern, we conducted additional full-image PSNR evaluations against TokenGS. GeoGSplat consistently achieves the highest PSNR across all target-context distances, with its advantage over TokenGS generally increasing as viewpoint changes become larger.

DistanceGeoGSplat PSNR ↑TokenGS PSNR ↑ΔPSNR ↑Invisible ratio
0.0025.3125.080.230.00%
0.1124.6324.500.132.58%
0.2222.6121.840.777.98%
0.3321.9721.010.9610.88%
0.4421.6420.281.3613.99%
0.5620.9819.621.3617.14%
0.6720.3418.931.4119.27%
0.7819.5117.871.6422.52%
0.8918.5616.651.9125.03%
1.0014.7312.702.0341.51%

Fig. 5 excludes unobserved regions to isolate reconstruction quality. Feed-forward 3D Gaussian Splatting is a reconstruction framework rather than a generative model, aiming to accurately reconstruct regions supported by input observations rather than generate unobserved content.

Our main NVS results in Table 1 already use full-image evaluation. GeoGSplat consistently outperforms TokenGS and AnySplat* on both DTU and NRGBD, further confirming that its advantages do not depend on visibility masking.

DatasetTokenGS PSNR ↑AnySplat* PSNR ↑GeoGSplat PSNR ↑
DTU17.7117.5420.48
NRGBD23.8220.5224.78

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive feedback and positive assessments of our work, particularly regarding the importance of the problem, the improvements in NVS and reconstruction, and the experimental evidence. We are encouraged by reviewers bBqh and Qp4v for acknowledging the effectiveness of our architecture and supporting ablation studies. To address the remaining concerns, we conducted additional experiments and summarize the key findings below.

1. Geometry Supervision (Reviewers bBqh, HHDw, Qp4v)

Simply adding geometry loss without a design that effectively leverages geometry is insufficient. To demonstrate this, we retrained TokenGS using its original novel-view photometric supervision and visibility loss, both with and without additional geometry loss. Although geometry loss improves TokenGS, it still substantially underperforms GeoGSplat, particularly in geometric accuracy.

MethodDTU PSNR ↑DTU SSIM ↑DTU LPIPS ↓DTU CD ↓NRGBD PSNR ↑NRGBD SSIM ↑NRGBD LPIPS ↓NRGBD CD ↓
TokenGS15.240.5930.3170.46023.240.7780.1760.191
TokenGS + geometry loss16.370.6080.3140.03824.430.8070.1530.029
GeoGSplat20.480.7710.1370.01824.780.8470.1200.017

GeoGSplat does not require perfectly accurate or complete GT depth. Although geometry supervision is important for learning explicit intermediate geometry, additional experiments using our lightweight model show that noisy or missing depth causes PSNR changes of at most 0.05 dB, demonstrating stable NVS performance under imperfect depth supervision. Specifically, we add Gaussian noise (σ = 5%) to randomly selected 10%–90% of depth pixels or remove depth supervision from one of the four input views.

Training settingDTU PSNR ↑DTU CD ↓NRGBD PSNR ↑NRGBD CD ↓
Original19.790.01923.700.019
Noisy depth19.740.03023.750.025
Missing depth19.760.01923.670.020

2. Computational Efficiency (Reviewers bBqh, HHDw, Qp4v)

The efficiency emphasized in our paper is the representation efficiency of non-pixel-aligned methods over pixel-aligned methods. Pixel-aligned methods produce Gaussians in proportion to the input resolution and number of views, resulting in substantial redundancy, whereas non-pixel-aligned methods maintain a fixed Gaussian budget. To verify whether this representation efficiency translates into practical computational benefits while achieving superior NVS and reconstruction quality, we conducted additional computational profiling against AnySplat across 12–64 input views. At 64 views, GeoGSplat uses 108× fewer Gaussians, achieves 3.08× faster inference and 4.22× higher rendering throughput, and reduces peak GPU memory by 77.2%.

3. Generalization Under Large Viewpoint Changes (Reviewers bBqh, HHDw)

GeoGSplat's advantages under large viewpoint changes remain valid under full-image evaluation, without relying on visibility masking. Additional full-image PSNR/LPIPS evaluations against TokenGS and AnySplat* confirm that GeoGSplat achieves the highest full-image PSNR across all distances. At the largest viewpoint change, GeoGSplat outperforms TokenGS by 2.03 dB and AnySplat* by 2.76 dB, even when previously unobserved regions are included.

Fig. 5 excludes unobserved regions to isolate reconstruction quality. Feed-forward 3D Gaussian Splatting is a reconstruction framework rather than a generative model, aiming to accurately reconstruct regions supported by input observations rather than generate unobserved content.

4. Inference Without GT Scene Scale or GT Camera Poses (Reviewers bBqh, HHDw, Qp4v)

GeoGSplat does not require GT scene scale or GT camera poses at inference time. GT scene scale is used only for evaluation and applied to all baselines to ensure fair comparisons and optimal performance. At inference time, GeoGSplat only requires approximate scene-scale normalization, similar to TokenGS, and can use estimated camera poses instead of GT poses. Our additional experiments demonstrate stable NVS performance under both settings, supporting the practical applicability of GeoGSplat.


Detailed experimental results and clarifications are provided in our individual reviewer responses. We will revise the manuscript to incorporate these clarifications and additional experimental results. We thank the reviewers and Area Chair again for their consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.